Unbalanced panel data: FE and robust SE's

[lauren29]

Hello,

I am working on an unbalanced panel data (T=17 and N=225). A sample would look like ;

FirmID Year Industry Y Var1 Var2
xyz 1997 Automobile 10 12 6
xyz 1998 Automobile 11 13 1
xyz 1999 Automobile 19 4 8
zzz 2003 Utilities 5 3 7
zzz 2004 Utilities 7 9 4



I have basically followed the guidance from this document : http://www.econ.canterbury.ac.nz/person ... _Week3.pdf
Despite the fact that T<N I estimated a cross sectional Fixed effects model for theoretical reasons. My panel being unbalanced I am unable to use the SUR options when selecting the GLS weighting. As far as Cross Section Weights are concerned, I get the following error message "Positive or non negative argument to function expected in computation of group weight (variance)". So I am not using GLS weights,
Therefore, I am mainly preoccupied with the standard erros and covraiance adjustements. I am not sure of the option I should pick among White cross-section, white period or diagonal. After some reading, I understood that :

-White cross-section method assumes that the errors are contemporaneously correlated (period clustered). It is cross-sectional dependant robust
-White period assumes that the errors for a cross section are heterosckedastic and serially correlated (cross section cluster). The estimator is designed to accommodate arbitrary heteroskedasticity and within cross section serial correlation
-The white diagonal method is robust to observation specific heteroskedasticity in the disturbances but not to correlation between residuals for different observations.

So far I have noticed that using the white diagonal leads to way higher SE's, which in terms of inference lead to accept the significance of almost all explanatory variables of my database.

I would pick white cross section (I believe it makes more sense considering my data) and in ordrer to control for serial correlation (having so far a low DW stat), I would add the lagged dependent variable to my specification.

1/ Is that a good approach?

2/ Is there any way in EViews to specificy that the cluster I have in mind is "Industry" rather than Firm ID when computing the coeff covariance method without restructuring the whole database ? Alternatively I tried to include dummy for Industries in the equation (10 industries, so I added 10 - 1=9 dummies but I get the "Near Singular Matrix" error)

Thanks a lot for your help :D

PS/ I already went through most of the topics in this forum about the subject ( such as http://forums.eviews.com/viewtopic.php?t=362&f=4) but I am still confused

[EViews Glenn]

Please post the workfile.

[lauren29]

Done ;)

[EViews Glenn]

1. You can't estimate with cross-section weights since you have some cross-sections with only a single observation. In this case, the variance of the residuals within cross-section is 0.

2. You are going to have asymptotic bias issues if you estimate with cross-section fixed effects and a lagged endogenous variable with small T. Here's a quick summary

http://faculty.washington.edu/ezivot/ec ... slides.pdf

3. You'll have to restructure to cluster by industry.

4. Unless firms move between industries, you can't have both firm fixed effects and industry effects. They'll be colinear.

[lauren29]

Thank you very much EViews Glenn for that clarification.So I think I will drop the fixed effects and rather include simple dummy variables for each industry. Still need to tackle autocorrelation though, I hope adding an AR(1) term and a robust covariance correction will do it.
Thanks!

[EViews Glenn]

Note that the AR term does implicitly add a lagged endogenous regressor.

[adnma]

I want to run a fixed effects model with HAC standard errors enabled....In cross section equation estimation window to enable Rob. SEs we go to options>and than choose EITHER "White" or "HAC Newey west(for time series data)"

How do I enable the same for panel data...as in equation estimation window, under options and panel options this feature is missing!

Please advice me in this regard.

[EViews Gareth]

The panel options tab has a Coef Covariance Method dropdown...